
OpenAI Releases GPT-6.2 Astra, the First Public Model It Rates as Able to Speed Up AI Research
- Transcript
AI Newspaper
AI Newspaper is your daily podcast for understanding what’s happening in artificial intelligence and why it matters.
Each episode focuses on one important AI story, development, or topic, breaking down what happened, the context behind it, and what it could mean next. From new models and breakthrough research to AI tools, industry shifts, regulation, and the companies shaping the future, AI Newspaper goes beyond the headline to give you the bigger picture.
New episodes daily. Stay informed about AI without having to follow the news all day.

OpenAI's new GPT-6.2 Astra is the first model the company has released publicly that meets the "High" threshold for AI self-improvement under its Preparedness Framework, meaning OpenAI judges it able to meaningfully speed up AI research. We look at what the GPT-6.2 system card says about that rating, the Critical cybersecurity rating both Astra and Sol carry, the tighter safeguards and the new Control Room system that can halt model workloads, and how the launch pairs those warnings with lower prices and new highs on professional-work benchmarks. Visit ainewspaper.org for today's stories and free email alerts.

The Federal Reserve's latest Financial Stability Report found that 87% of the market contacts it surveyed now name artificial intelligence as a top risk to the US financial system, up from 50% in the spring and well ahead of every other concern. We break down the three AI worries respondents flagged: stretched valuations and the chance of a correction, cyber risk from state actors or frontier models themselves, and increasingly complex AI financing. We also look at the Fed's numbers on the data center buildout, including more than $680 billion in advanced computing-related debt issued and over $1.1 trillion outstanding. Visit ainewspaper.org for today's stories and free email alerts.

Four nonprofits, led by Public Citizen, have asked a San Francisco court to stop Anthropic, Google, Meta, OpenAI and SpaceX's AI unit from building or releasing any AI more capable than what they had on November 12, 2026. The suit, filed under California's Unfair Competition Law, argues frontier AI is a public nuisance and leans on the companies' own warnings about bioweapons, cyberattacks and loss of control. We walk through what the groups are asking for, including a ban on AI autonomously directing AI research, the examples of AI agents already causing harm cited in the complaint, and the legal hurdles the case faces before any judge could order a freeze. Visit ainewspaper.org for today's stories and free email alerts.

Jensen Huang has resigned as Nvidia's president, CEO and board member, effective immediately, after reports of two extramarital affairs, including one with a woman accused of spying for China. CFO Colette Kress is now interim CEO while keeping her finance role, and longtime director Harvey Jones becomes the company's first-ever board chair as Nvidia searches for a permanent successor. We look at how the market reacted, with Nvidia shares diving as much as 16% before closing down 4.3% while AMD and Broadcom rose about 6%, what the board and Kress told employees, and what Huang's exit means for the company at the center of the AI boom, with third-quarter earnings still set for November 17. Visit ainewspaper.org for today's stories and free email alerts.

Anthropic has launched the roadshow for its initial public offering, aiming to sell 825 million shares at $70 to $78 each on the Nasdaq under the ticker ANT. At the top of the range it would raise about $64 billion and be valued at up to $1.4 trillion, with chipmakers and cloud giants adding a $26 billion private placement. We look at what the prospectus reveals, from roughly $81 billion in run-rate revenue to nearly $901 billion in long-term compute commitments and an unusually long list of AI safety risks, why the valuation came in below earlier hopes of $2 trillion, and why a Federal Reserve official is warning of a large market correction if AI returns disappoint. Pricing is expected November 18, with trading on November 19. Visit ainewspaper.org for today's stories and free email alerts.

Google DeepMind says AI agents have mathematically proved key security properties of the actual compiled code behind the Signal protocol, the encryption that protects private conversations in Signal and Google Messages. Using a new framework called AlphaProof Code, agents running on Gemini Carbon generated about 6.5 million lines of machine-checked Lean proof in four weeks, covering Signal's post-quantum handshake and Triple Ratchet. We look at what was proved, how humans and agents split the work, what the proof does not cover, and why Signal's CTO calls it a major breakthrough, on the same day Gemini 4 Carbon became generally available. Visit ainewspaper.org for today's stories and free email alerts.

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, and its headline experiment is striking: starting from an empty folder, Fable 5.2 coordinated more than 1,000 subagents for eight days and produced a working operating system that boots on a laptop. A separate team of 1,000 agents formalized Fermat's Last Theorem in about 30 hours. We look at what Anthropic says the new models can do, how pricing and access differ between Fable and Mythos, and why the system card says the models voice suspicion they are being tested more often than any recent Claude model. Visit ainewspaper.org for today's stories and free email alerts.

Anthropic has reportedly agreed to pay about $3.4 billion for commitments covering 1.7 gigawatts of heavy-duty gas-fired generators, taking over Caterpillar generator orders that Texas-based OnePWR Solutions placed about two years ago, along with the engineering and construction capability to install them. It is described as the largest known single purchase of power-generation equipment by an AI lab. Anthropic did not buy OnePWR, and neither company has commented. We look at why turbines and generators have become as scarce as chips, what Anthropic's own filings say about on-site gas power, and how the frontier race is pulling AI companies directly into the energy business. Visit ainewspaper.org for today's stories and free email alerts.

Google DeepMind published two working papers in which human-guided AI agents sharpened attacks on legacy 1024-bit RSA. In a measured reproduction of a 2026 record computation, agents forged an RSA signature root for 391 CPU core-years versus 1,380 originally, about 3.5 times less, in roughly three days of full-scale work instead of five months. The ~1,600x signature-forgery speedup and ~30x "fan factory" batch-factoring gains are theoretical, leading-order estimates; no fan factory has been built. Both papers say modern 2048-bit and larger RSA is unaffected. We look at what it means when encryption can no longer count on hardness fading at the old human pace, and why leftover 1024-bit keys should be retired now. Sources: DeepMind papers (storage.googleapis.com/deepmind-media), alfw.ai, Tao Nhu Blog. Visit ainewspaper.org for today's stories and free email alerts.

A research system at the startup Physical Superintelligence (PSI), running instances of a non-public Anthropic model, has produced a proof of the Aaronson-Ambainis conjecture — a question about the limits of quantum speedups that had been open since 2008. The system had a complete proof on the third day of its run, and the Lean proof assistant accepted it. The open question now is how fast humans can check and understand it. PSI announced the result on November 6, 2026. The conjecture, posed by Scott Aaronson and Andris Ambainis, is a statement about bounded low-degree polynomials: it says some input bit must have meaningful influence on the polynomial's value. Aaronson and Ambainis showed that if it holds, quantum computers can be at most polynomially — not exponentially — faster than ordinary computers at answering yes-or-no questions about random inputs that lack special structure. PSI's proof follows a plan by Shachar Lovett and Jiapeng Zhang, replacing the one inequality they had left as a conjecture with a new inequality proved via a determinant-based pairing that strengthens David Reimer's Butterfly Lemma. After the first proof, PSI set the system on related problems and it improved the result twice; Lean accepted the final version. The paper and a Lean 4 formalization are public. The repository says the main theorems have no placeholders or added axioms and were re-checked by two independent implementations of the Lean kernel, and that the proof, paper and formalization were produced by AI models. Machine checks have limits: a human still has to confirm the formal statement matches the printed conjecture, and the human-readable paper is not itself machine-checked. Stanford researchers Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan, who settled a special case in August, called it "a huge result" and are writing a shorter exposition for experts. PSI's David Turturean, who built the system, reviewed the draft and is responsible for the paper's correctness. Business Insider reports it is not clear which unreleased Anthropic model PSI used, and Anthropic did not respond to a request for comment. Visit ainewspaper.org for today’s stories and free email alerts.

OpenAI published research on an autonomous vulnerability discovery, exploitation, and patching pipeline that, in a mid-October 2026 run against a stock Linux desktop OS in isolated VMs, produced 948 working local privilege-escalation exploits that could let an unprivileged local user gain root. On November 4, 2026, OpenAI described the pipeline in a post titled "A thousand fewer paths to root." Powered by an upcoming cyber model, agents analyzed 2,308 candidate kernel findings, dynamically reproduced 1,724 vulnerabilities, and generated 948 working LPE exploits plus 291 information-leak proofs of concept. Total API cost: about $1.21 million, or roughly $1,279 per LPE. Agents reportedly used at least 10 novel Linux exploitation techniques. Human researchers review findings before reporting; hundreds of critical reports and patches have already gone upstream with Patch the Planet partners. When the campaign finished, 61 LPEs were already patched upstream, 43 had staging patches, and 299 were being discussed publicly. A July GPT-5.6 Sol run had found 166 LPEs; this run found more than five times as many. Attacks still need an existing foothold, and the work was done in a research setting only. The episode looks at what it means when finding bugs is no longer the scarce resource — reviewing and shipping patches before attackers catch up is. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

Independent research group METR says Anthropic is likely improving AI model capabilities on well-specified tasks at around twice the historical rate — largely because AI itself is doing much of the R&D. The catch: Anthropic redacted most of the evidence outsiders would need to check the claim. On November 5, 2026, METR published a high-access assessment of AI R&D acceleration at Anthropic covering August 12 through October 29, 2026. METR’s subjective conclusion: as of October 2026 Anthropic is likely improving capabilities on well-specified tasks at around twice the rate from March 2024–October 2025, with high uncertainty (METR would not be surprised by roughly 1.3x–3x). Measured via AECI (Anthropic’s Epoch Capabilities Index variant), internal model benchmark scores have grown at an estimated 1.9x the pre-October 2025 rate [95% CI 1.5–2.8x] — roughly two years of historical progress in one year. METR is confident the acceleration was not from faster training-compute scaling. It guesses a majority (perhaps about 80%) of the current October acceleration is due to AI used in R&D — mainly AI assistance rather than fully autonomous contributions, with “signs of life” for more autonomy soon — and reports no clear bottlenecks that would block further acceleration in the next 12 months. Caveat front and center: while METR could share its highest-level risk conclusions, Anthropic redacted a majority of the evidence METR found most informative in the roughly 100-page assessment, describing redactions as aimed at removing confidential information that could accelerate other AI developers. AP and The Verge also covered the report. Visit ainewspaper.org for today’s stories and free email alerts.

The New York Times has sued Brave Software, the company behind the Brave browser and Brave Search, accusing it of acting like "a burglar" that takes publishers' journalism and an AI "fence" that sells it to others. The suit was filed on Monday, November 2, in federal court in Manhattan, and the same day the Daily News, the Chicago Tribune and the Orlando Sentinel filed a similar case. The Times says Brave built a search index of more than 30 billion web pages with a crawler it has never named, so publishers have no way to block it. It says Brave then used that index to power its own AI answers in Ask Brave, and sold access to it through the Brave Search API to AI companies the complaint names as Anthropic, Perplexity, Mistral, xAI and Cohere. In one test described in the complaint, Brave's LLM Context endpoint returned 1,291 of the 1,295 words in a Times article. The suits also target Brave's "Paywall blocker" browser feature and its practice of blocking publishers' ads while selling its own. The Times says it sent cease-and-desist letters to Brave's CEO, Brendan Eich, in August 2025 and April 2026, and that both were ignored. It is asking for damages, the profits Brave made, and a permanent court order. Brave did not immediately respond to requests for comment. The episode looks at why publishers are taking the fight over AI crawlers to court, and why the search indexes that quietly supply AI companies are becoming the next target. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

Abu Dhabi has signed four agreements with OpenAI as part of its goal to become what it calls the world's first AI-native government. The deals were signed on Monday, November 2, after Crown Prince Sheikh Khaled bin Mohamed bin Zayed met OpenAI CEO Sam Altman, who is in Abu Dhabi for the ADIPEC energy show. The centerpiece is a Machine-Readable Government initiative: a defined set of government policies and regulations is being turned into a form AI systems can read, so AI can help officials with routine compliance checks such as issuing permits and certificates. Each assessment is meant to be traceable to the rule behind it, and officials keep oversight and responsibility for final decisions. Led by the Department of Government Enablement, the first pilot is already running with the Abu Dhabi Agriculture and Food Safety Authority on food imports, with the AI running on OpenAI's GPT models on GPUs inside UAE data centres. In energy, ADNOC, G42's subsidiary Inception42 and OpenAI signed a deal that makes ADNOC the first energy company to deploy OpenAI's Frontier technology across its operations. Mubadala, Inception42 and OpenAI will apply AI to investment analysis, opportunity identification and risk assessment, and Inception42 and OpenAI signed a five-year agreement to spread AI across government, enterprise and education in the UAE. The episode also covers the long ties between OpenAI and Abu Dhabi, from Stargate UAE to MGX's role in OpenAI's latest funding round, and looks at how much of the state's everyday judgment moves into an AI company's models, and how officials keep the final say. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

An AI agent has chosen a gene therapy and injected it into the brain of a live mouse. Gabriel Ocana-Santero, CEO and cofounder of the San Francisco biotech startup Exin Therapeutics, says his team gave a Claude Sonnet 5 agent a spreadsheet of 19 gene therapy options, a mouse model of epilepsy, a motorized surgery rig with cameras, and one goal: reduce seizures. The agent picked a therapy meant to boost the brain's inhibitory signaling, chose the lateral entorhinal cortex as the target, worked out injection coordinates from three scientific papers, and then drove the needle and injected 250 nanolitres while watching the site through a camera for bleeding. A veterinary surgeon opened the skull, a licensed veterinarian approved every irreversible step and held a physical emergency stop, and the agent was barred from overriding fixed safety limits. Midway through, after a connection dropped, the agent refused to move the needle because its target fell outside the rig's range, and it turned out to be right: earlier manual adjustments had thrown off the controller. Anthropic's more capable Claude Opus 5.5 and Opus 5 models both declined to carry out the live procedure. The mouse is recovering, but the effect on seizures has not yet been measured, the results have not been peer-reviewed, and a technical report is still to come. The episode looks at what changes when AI agents are connected directly to lab equipment, and who is in control when a model can pick the treatment and also say no. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

Alphabet lost about $500 billion in stock market value on Thursday, October 29, after its third-quarter results showed Google's own AI search features cutting into the ad business that funds its AI spending. Shares fell about 13%, their worst day since Google went public in 2004 and the second-largest one-day loss of value in U.S. market history, behind only Nvidia's $589 billion drop in January 2025. Search and other revenue grew 12% to $63.35 billion but missed estimates for the second straight quarter, even as Google said search queries grew at their fastest pace in five years. Search growth has slowed from 19% in the first quarter to 17% in the second and 12% in the third. Chief Business Officer Philipp Schindler said AI Mode has lower monetization than traditional search and that Google is now scaling ads in AI Overviews and AI Mode. Alphabet reported negative free cash flow of $9.07 billion, against about $2.7 billion expected, kept its 2026 capital spending forecast at $195 billion to $205 billion, and has more than $1 trillion in multiyear commitments for data centers, computing and energy. Total revenue rose 25% to $127.6 billion, Google Cloud revenue more than doubled to nearly $31 billion, and the Gemini app now has more than 1 billion users. The episode also covers the results from Microsoft and Amazon, and what Wall Street is asking about where AI money will ultimately be made. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

Canada's government plans to invest up to C$1.6 billion for an equity stake in Toronto-based AI company Cohere. Prime Minister Mark Carney announced the commercial investment on Thursday, October 29, in Montreal, where he was hosting the Canada-European Union Summit. The money is part of Cohere's planned C$7 billion financing, led by Germany's Schwarz Group, which would value the company at about C$36 billion and would be the largest funding round ever by a Canadian startup. Nvidia, Bell, Bullhound Capital, Inovia Capital and Prosperity7 Ventures are also taking part. In return, Cohere has committed to keep its Canadian incorporation and headquarters, create at least 1,000 new technical jobs in Canada by 2030, put at least C$1.7 billion into computing infrastructure in Canada by 2030, fund research partnerships and keep working with the Canadian AI Safety Institute. The investment comes through the Canada Development Investment Corporation and will be paid in stages tied to jobs and investment targets, with regular progress reports. The episode also covers Cohere's merger with Germany's Aleph Alpha, its revenue growth, and the competition it faces from much larger AI companies in a costly industry where profits have been hard to find. Cohere has completed only an initial close of the round and has not said when it will close. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

The Dutch cabinet expects central government to have about 25,000 fewer full-time jobs by 2035, roughly 20% of the civil service, and says AI will help it get there. State secretary Eric van der Burg set out the plan in a letter to parliament on Monday, October 26. The cuts should save €900 million in 2027, rising to €3.5 billion a year from 2035. Police and the armed forces are exempt, and some ministries, inspectorates and delivery agencies could see staff cut by 45% or even 60%. The plan follows rapid growth: central government reached a record of about 160,000 full-time jobs at the end of 2025, 41% more than in 2018. The episode also covers the plan to scrap or simplify at least 500 rules a year, one desk for every two civil servants, a preference for Dutch and European suppliers, and why economists and an OECD report question whether government AI is ready to deliver. The split of cuts by ministry and the cabinet's AI goals and savings plans have not been published yet; they are due in 2027, with AI plans in the first quarter and the ministry split in the spring budget update. Visit ainewspaper.org to see today's stories and sign up for free email alerts.

Anthropic, the maker of Claude, has publicly filed to go public. Its Form S-1, filed with the US Securities and Exchange Commission on October 26, says it plans to list Class A shares on the Nasdaq under the ticker "ANT". The number of shares and the price range have not been set yet. The filing shows first-half 2026 revenue of $16.3 billion, up from $1.2 billion a year earlier, and an operating loss that narrowed to $693 million. It also shows an $80.7 billion net loss, most of it a non-cash accounting charge tied to convertible notes. The episode covers Anthropic's enterprise customer numbers and roughly 900% net revenue retention, its reliance on its largest customers and on the Amazon, Google and Microsoft cloud marketplaces, and its unusual governance: a public benefit corporation where a Long-Term Benefit Trust elects most of the board and a single founders' share carries 50.1% of the vote. All figures come from a preliminary prospectus that may still change before the offering.

Apollo Global Management chief economist Torsten Slok argues that, so far, AI has delivered no visible financial payoff to most of corporate America. In his Daily Spark notes, Slok compares the "S&P 493", the S&P 500 minus Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla, with the biggest tech companies: revenue per worker and profit margins are flat for the 493 while both are rising for the Magnificent Seven. He also finds no sign of AI in the productivity data, with the San Francisco Fed's utilization-adjusted total factor productivity slightly below zero and gains in output per hour explained by capital deepening. This is one economist's analysis of public data in an opinion note, not a peer-reviewed study. The episode covers why mainstream companies are starting to treat AI token spending as a cost, Big Tech's roughly $650 billion in AI spending this year, and why Microsoft, Alphabet and Amazon earnings this week matter for the debate.